GPU Seismic Modeling Subsets for Memory and Bandwidth Reduction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current seismic modeling on GPUs faces challenges with limited memory requirements, high GPU-CPU transfer bandwidth, and prolonged setup times, which hinder accurate and efficient processing of complex seismic data.
Innovation Solution
A method and system that divide seismic data into subsets allocated on GPU memory, allowing simultaneous computation and data transfer, with a small portion of data stored on GPU for computation and the rest on CPU or HDD, enabling efficient data compression and overlapping computation with data transfer, and reversing loop direction to minimize setup overhead.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Speed
If all seismic data is loaded into GPU memory for computation, then computation speed is improved, but GPU memory requirements increase significantly
Solution Approach 1:
The patent divides the discretized model into multiple subsets that are allocated in GPU memory. Only the currently needed subset is loaded into GPU memory for computation, while other subsets remain on CPU or disk. This segmentation allows the system to process large seismic datasets without requiring the entire dataset to be simultaneously present in GPU memory, thus resolving the contradiction between computation speed and memory requirements.
2Adaptability or versatility
If data is transferred between CPU and GPU during computation, then data accessibility is improved, but transfer bandwidth requirements increase
Solution Approach 1:
The patent pre-loads the currently needed subset data into GPU memory before computation begins. By preparing the data in advance and having it ready in GPU memory when needed, the system minimizes the frequency and volume of data transfers during the computation process. This preliminary action reduces the overall transfer bandwidth requirements while maintaining data accessibility.
3Productivity
If data is pre-processed and loaded into GPU memory before computation, then computation efficiency is improved, but setup time increases
Solution Approach 1:
The patent loads only the currently needed subset of data into GPU memory rather than pre-processing and loading all data. This partial action approach maintains computation efficiency by having the necessary data ready in GPU memory while significantly reducing setup time compared to loading the entire dataset. The system performs just enough data preparation to enable efficient computation without the overhead of complete pre-processing.
Data Source
AI summary
A method for seismic modeling on a GPU includes creating a discretized model containing a volume enclosed by boundaries with arbitrary boundary conditions, dividing the discretized model into subsets, transferring a data to the GPU, initializing the data stored in a selected number of the subsets on the GPU, performing computation related with a number of time steps for the selected number of the subsets on the GPU, and substantially simultaneously computing, off-loading a finished subsets data from the GPU, and up-loading a new subsets data to the GPU. The provided method and corresponding system can reduce the amount of data stored in GPU memory by an order of magnitude, minimize the GPU device-to-host transfer bandwidth requirements, and reduce the GPU setup time.